workflow orchestration

Production-grade autonomous AI workflows face significant engineering challenges in reliability, observability, maintainability, and security governance. Method: We propose a structured, full-lifecycle methodology comprising a multi-agent architecture with collaborative reasoning, tool augmentation, and dynamic orchestration—integrated with the Model Context Protocol (MCP), deterministic orchestration, pure function invocation, containerized deployment, and modular tool integration. We further define nine core engineering practices, including tool-first design, single-responsibility agents, externalized prompt management, and model-federation-driven responsible AI design. Contribution/Results: This work establishes the first systematic engineering paradigm for Agentic AI productionization, markedly improving system simplicity, observability, and governability. Empirical validation via a multimodal news analysis–media generation use case demonstrates robustness and scalability. The methodology provides a reusable framework and practical benchmark for industrial-scale autonomous AI systems.

This study addresses the challenge of automating workflows in complex industries—such as logistics, healthcare, and construction—where processes are fragmented across heterogeneous tools and involve multi-party collaboration. The work proposes orchestration as a core abstraction to enable effective automation by dynamically coordinating multi-step tasks, enforcing domain-specific constraints, managing human approvals, and integrating legacy systems. It introduces the novel concept of “orchestration bottlenecks” and develops a theoretical framework that unifies multi-agent systems, workflow modeling, constraint reasoning, and human–AI collaboration, while exposing critical gaps in current multi-agent approaches at the orchestration level. Based on distinct sources of operational friction across domains, the paper advocates for targeted architectural safeguards—such as constraint enforcement or explainability—and phased implementation strategies to provide actionable pathways for automation in complex operational environments.

This work addresses the absence of a unified conceptual framework for describing the autonomy of AI agents and the allocation of decision-making authority in contemporary CI/CD pipelines. It introduces the notion of “authority transfer” to systematically delineate the boundaries of agent autonomy, distinguishing between decision rights in the data plane and the control plane, and identifies governance of the control plane as a critical research direction. Through architectural abstraction, pattern identification, and governance mechanism design—supported by prototype implementation and analysis of industrial platforms—the study reveals three prevalent patterns: constrained autonomy, externally dominated governance, and delayed evaluation. These findings establish a theoretical foundation and outline a research agenda for developing safe, controllable, and highly autonomous CI/CD systems.

This work addresses the current lack of open-source infrastructure capable of efficiently training and evaluating large-scale agents on complex tasks such as software engineering and computer operation. To this end, we propose a three-service decoupled architecture tailored for agent-environment interaction workloads, which separates the system into three independent services—model, agent, and environment—enabling fine-grained task scheduling, dynamic resource allocation, and unified interface communication. This design allows each component to scale independently and configure resources flexibly, significantly improving training efficiency and resource utilization. Experimental results demonstrate that the system can stably support tens of thousands of concurrent agent tasks, thereby filling a critical gap in infrastructure for large-scale agent training.

To address the lack of formal modeling support for human–AI collaborative workflows, this paper proposes a BPMN extension tailored for hybrid collaboration between humans and large language model (LLM)-based agents. We introduce the first BPMN metamodel and corresponding graphical notation explicitly supporting human–agent coordination, formally capturing task responsibility assignment, decision logic ownership, and execution strategy binding—thereby bridging a critical expressiveness gap in existing business process modeling languages for mixed-agent scenarios. Furthermore, we design a domain-specific language (DSL) and implement an open-source web-based modeling tool using TypeScript and React, enabling visual specification and formal verification of collaborative processes. Empirical evaluation through case studies demonstrates that the approach achieves strong expressive power, operational feasibility, and practical applicability in real-world human–LLM workflow engineering.

This work addresses the challenges of semantic drift and poor auditability in multi-step AI agent automation, which often arise from error propagation and sensitivity to prompt variations. The authors propose a framework that treats a visual workflow graph as the single source of truth, leveraging formal semantic definitions at compile time to specify data scopes, execution logic, and monitoring policies, while employing contract-based verification—via preconditions and postconditions—to ensure component correctness. At runtime, the system supports replay, retry, and auditing through persistent event logging and explicitly isolates risk using lane-based trust boundaries. Evaluated across three clinical institutions over 8,728 executions, the prototype achieved a 97.08% completion rate, with most failures attributable to external system integration issues.

This work addresses the limitations of existing human-in-the-loop (HITL) mechanisms in intelligent agent workflows, which are often tightly coupled with application logic, resulting in poor reusability, weak consistency, and limited scalability. To overcome these challenges, the paper proposes a decoupled HITL system architecture that abstracts human oversight into an independent component. By introducing explicit interfaces and a structured execution model, the approach cleanly separates human–machine interaction from business logic. Furthermore, it introduces a novel four-dimensional framework—comprising intervention conditions, role resolution, interaction semantics, and communication channels—to enable context-aware, controllable human intervention. This design achieves, for the first time, protocol-level reusability of HITL mechanisms, supporting consistent and scalable autonomy governance in multi-agent environments and laying a foundational infrastructure for system-level human–agent collaboration.

Current AI agents predominantly rely on ad hoc, on-the-fly generation strategies and lack the reliability safeguards inherent in established software engineering practices, rendering them ill-suited for high-stakes scenarios demanding stringent safety and robustness. This work proposes a novel “AI workflow store” paradigm that systematically integrates software engineering principles—such as iterative design, rigorous testing, and adversarial evaluation—into AI agent architectures. By encapsulating reusable workflows, enforcing deterministic execution constraints, and adopting phased deployment strategies, the framework constructs a hardened library of high-assurance AI workflows. Empirical results demonstrate that this approach significantly outperforms conventional just-in-time synthesis methods in mission-critical tasks, achieving markedly enhanced safety and robustness without sacrificing flexibility.

This study addresses the challenges of high latency, unstable concurrency, and security risks faced by large language model (LLM) agents in automating asset lifecycle management within Industry 4.0. The authors propose a Plan-then-Execute architecture that generates verifiable workflow graphs and integrates a topology-aware parallel scheduling mechanism to enable controlled inference overlap while ensuring functional correctness and security. Key technical contributions include topological-sort-based multi-agent scheduling, structured context pruning, dependency-aware concurrency control, and graceful degradation under fault injection. Evaluated on the AssetOpsBench benchmark, the system reduces median end-to-end latency by 1.6× (up to 1.8× for highly parallel tasks) and cuts inference overhead by approximately 30% through context pruning, all while maintaining stable task completion rates and output quality.

This work addresses the safety risks in highly flexible multi-agent manufacturing systems driven by large language models (LLMs), where frequent system reconfigurations can lead to hazardous behaviors due to the absence of formal safety guarantees. To mitigate this issue, the paper proposes a novel framework that integrates LLMs with formal verification by combining temporal logic and discrete event systems to validate LLM-generated task allocation plans prior to execution. This approach uniquely synergizes the adaptive planning capabilities of LLMs with logic-based safety assurance mechanisms. The framework has been validated in a multi-robot assembly scenario, demonstrating its ability to proactively identify and rectify unsafe behaviors, thereby ensuring the safety and reliability of the manufacturing process.